This section will cover the basic CRUD operations commonly performed on a relational database:


There is also a large collection of example queries taken from the Postgresql Exercises website. Examples are listed on the query examples document.

Creating a new record

You can use Model.create() to create a new model instance. This method accepts keyword arguments, where the keys correspond to the names of the model’s fields. A new instance is returned and a row is added to the table.

>>> User.create(username='Charlie')
<__main__.User object at 0x2529350>

This will INSERT a new row into the database. The primary key will automatically be retrieved and stored on the model instance.

Alternatively, you can build up a model instance programmatically and then call save():

>>> user = User(username='Charlie')
>>> user.save()  # save() returns the number of rows modified.
>>> user.id
>>> huey = User()
>>> huey.username = 'Huey'
>>> huey.save()
>>> huey.id

When a model has a foreign key, you can directly assign a model instance to the foreign key field when creating a new record.

>>> tweet = Tweet.create(user=huey, message='Hello!')

You can also use the value of the related object’s primary key:

>>> tweet = Tweet.create(user=2, message='Hello again!')

If you simply wish to insert data and do not need to create a model instance, you can use Model.insert():

>>> User.insert(username='Mickey').execute()

After executing the insert query, the primary key of the new row is returned.


There are several ways you can speed up bulk insert operations. Check out the Bulk inserts recipe section for more information.

Bulk inserts

There are a couple of ways you can load lots of data quickly. The naive approach is to simply call Model.create() in a loop:

data_source = [
    {'field1': 'val1-1', 'field2': 'val1-2'},
    {'field1': 'val2-1', 'field2': 'val2-2'},
    # ...

for data_dict in data_source:

The above approach is slow for a couple of reasons:

  1. If you are not wrapping the loop in a transaction then each call to create() happens in its own transaction. That is going to be really slow!

  2. There is a decent amount of Python logic getting in your way, and each InsertQuery must be generated and parsed into SQL.

  3. That’s a lot of data (in terms of raw bytes of SQL) you are sending to your database to parse.

  4. We are retrieving the last insert id, which causes an additional query to be executed in some cases.

You can get a significant speedup by simply wrapping this in a transaction with atomic().

# This is much faster.
with db.atomic():
    for data_dict in data_source:

The above code still suffers from points 2, 3 and 4. We can get another big boost by using insert_many(). This method accepts a list of tuples or dictionaries, and inserts multiple rows in a single query:

data_source = [
    {'field1': 'val1-1', 'field2': 'val1-2'},
    {'field1': 'val2-1', 'field2': 'val2-2'},
    # ...

# Fastest way to INSERT multiple rows.

The insert_many() method also accepts a list of row-tuples, provided you also specify the corresponding fields:

# We can INSERT tuples as well...
data = [('val1-1', 'val1-2'),
        ('val2-1', 'val2-2'),
        ('val3-1', 'val3-2')]

# But we need to indicate which fields the values correspond to.
MyModel.insert_many(data, fields=[MyModel.field1, MyModel.field2]).execute()

It is also a good practice to wrap the bulk insert in a transaction:

# You can, of course, wrap this in a transaction as well:
with db.atomic():
    MyModel.insert_many(data, fields=fields).execute()


SQLite users should be aware of some caveats when using bulk inserts. Specifically, your SQLite3 version must be or newer to take advantage of the bulk insert API. Additionally, by default SQLite limits the number of bound variables in a SQL query to 999 for SQLite versions prior to 3.32.0 (2020-05-22) and 32766 for SQLite versions after 3.32.0.

Inserting rows in batches

Depending on the number of rows in your data source, you may need to break it up into chunks. SQLite in particular typically has a limit of 999 or 32766 variables-per-query (batch size would then be 999 // row length or 32766 // row length).

You can write a loop to batch your data into chunks (in which case it is strongly recommended you use a transaction):

# Insert rows 100 at a time.
with db.atomic():
    for idx in range(0, len(data_source), 100):

Peewee comes with a chunked() helper function which you can use for efficiently chunking a generic iterable into a series of batch-sized iterables:

from peewee import chunked

# Insert rows 100 at a time.
with db.atomic():
    for batch in chunked(data_source, 100):


The Model.bulk_create() method behaves much like Model.insert_many(), but instead it accepts a list of unsaved model instances to insert, and it optionally accepts a batch-size parameter. To use the bulk_create() API:

# Read list of usernames from a file, for example.
with open('user_list.txt') as fh:
    # Create a list of unsaved User instances.
    users = [User(username=line.strip()) for line in fh.readlines()]

# Wrap the operation in a transaction and batch INSERT the users
# 100 at a time.
with db.atomic():
    User.bulk_create(users, batch_size=100)


If you are using Postgresql (which supports the RETURNING clause), then the previously-unsaved model instances will have their new primary key values automatically populated.

In addition, Peewee also offers Model.bulk_update(), which can efficiently update one or more columns on a list of models. For example:

# First, create 3 users with usernames u1, u2, u3.
u1, u2, u3 = [User.create(username='u%s' % i) for i in (1, 2, 3)]

# Now we'll modify the user instances.
u1.username = 'u1-x'
u2.username = 'u2-y'
u3.username = 'u3-z'

# Update all three users with a single UPDATE query.
User.bulk_update([u1, u2, u3], fields=[User.username])

This will result in executing the following SQL:

UPDATE "users" SET "username" = CASE "users"."id"
    WHEN 1 THEN "u1-x"
    WHEN 2 THEN "u2-y"
    WHEN 3 THEN "u3-z" END
WHERE "users"."id" IN (1, 2, 3);


For large lists of objects, you should specify a reasonable batch_size and wrap the call to bulk_update() with Database.atomic():

with database.atomic():
    User.bulk_update(list_of_users, fields=['username'], batch_size=50)


Model.bulk_update() may not be the most efficient method for updating large numbers of records. This functionality is implemented such that we create a “mapping” of primary key to corresponding field values for all rows being updated using a SQL CASE statement.

Alternatively, you can use the Database.batch_commit() helper to process chunks of rows inside batch-sized transactions. This method also provides a workaround for databases besides Postgresql, when the primary-key of the newly-created rows must be obtained.

# List of row data to insert.
row_data = [{'username': 'u1'}, {'username': 'u2'}, ...]

# Assume there are 789 items in row_data. The following code will result in
# 8 total transactions (7x100 rows + 1x89 rows).
for row in db.batch_commit(row_data, 100):

Bulk-loading from another table

If the data you would like to bulk load is stored in another table, you can also create INSERT queries whose source is a SELECT query. Use the Model.insert_from() method:

res = (TweetArchive
           Tweet.select(Tweet.user, Tweet.message),
           fields=[TweetArchive.user, TweetArchive.message])

The above query is equivalent to the following SQL:

INSERT INTO "tweet_archive" ("user_id", "message")
SELECT "user_id", "message" FROM "tweet";

Updating existing records

Once a model instance has a primary key, any subsequent call to save() will result in an UPDATE rather than another INSERT. The model’s primary key will not change:

>>> user.save()  # save() returns the number of rows modified.
>>> user.id
>>> user.save()
>>> user.id
>>> huey.save()
>>> huey.id

If you want to update multiple records, issue an UPDATE query. The following example will update all Tweet objects, marking them as published, if they were created before today. Model.update() accepts keyword arguments where the keys correspond to the model’s field names:

>>> today = datetime.today()
>>> query = Tweet.update(is_published=True).where(Tweet.creation_date < today)
>>> query.execute()  # Returns the number of rows that were updated.

For more information, see the documentation on Model.update(), Update and Model.bulk_update().


If you would like more information on performing atomic updates (such as incrementing the value of a column), check out the atomic update recipes.

Atomic updates

Peewee allows you to perform atomic updates. Let’s suppose we need to update some counters. The naive approach would be to write something like this:

>>> for stat in Stat.select().where(Stat.url == request.url):
...     stat.counter += 1
...     stat.save()

Do not do this! Not only is this slow, but it is also vulnerable to race conditions if multiple processes are updating the counter at the same time.

Instead, you can update the counters atomically using update():

>>> query = Stat.update(counter=Stat.counter + 1).where(Stat.url == request.url)
>>> query.execute()

You can make these update statements as complex as you like. Let’s give all our employees a bonus equal to their previous bonus plus 10% of their salary:

>>> query = Employee.update(bonus=(Employee.bonus + (Employee.salary * .1)))
>>> query.execute()  # Give everyone a bonus!

We can even use a subquery to update the value of a column. Suppose we had a denormalized column on the User model that stored the number of tweets a user had made, and we updated this value periodically. Here is how you might write such a query:

>>> subquery = Tweet.select(fn.COUNT(Tweet.id)).where(Tweet.user == User.id)
>>> update = User.update(num_tweets=subquery)
>>> update.execute()


Peewee provides support for varying types of upsert functionality. With SQLite prior to 3.24.0 and MySQL, Peewee offers the replace(), which allows you to insert a record or, in the event of a constraint violation, replace the existing record. For Sqlite 3.24+ and Postgres, peewee provides full support for ON CONFLICT queries.

Example of using replace() and on_conflict_replace():

class User(Model):
    username = TextField(unique=True)
    last_login = DateTimeField(null=True)

# Insert or update the user. The "last_login" value will be updated
# regardless of whether the user existed previously.
user_id = (User
           .replace(username='the-user', last_login=datetime.now())

# This query is equivalent:
user_id = (User
           .insert(username='the-user', last_login=datetime.now())


In addition to replace, SQLite, MySQL and Postgresql provide an ignore action (see: on_conflict_ignore()) if you simply wish to insert and ignore any potential constraint violation.

MySQL supports upsert via the ON DUPLICATE KEY UPDATE clause. For example:

class User(Model):
    username = TextField(unique=True)
    last_login = DateTimeField(null=True)
    login_count = IntegerField()

# Insert a new user.
User.create(username='huey', login_count=0)

# Simulate the user logging in. The login count and timestamp will be
# either created or updated correctly.
now = datetime.now()
rowid = (User
         .insert(username='huey', last_login=now, login_count=1)
             preserve=[User.last_login],  # Use the value we would have inserted.
             update={User.login_count: User.login_count + 1})

In the above example, we could safely invoke the upsert query as many times as we wanted. The login count will be incremented atomically, the last login column will be updated, and no duplicate rows will be created.

Postgresql and SQLite (3.24.0 and newer) provide a different syntax that allows for more granular control over which constraint violation should trigger the conflict resolution, and what values should be updated or preserved.

Example of using on_conflict() to perform a Postgresql-style upsert (or SQLite 3.24+):

class User(Model):
    username = TextField(unique=True)
    last_login = DateTimeField(null=True)
    login_count = IntegerField()

# Insert a new user.
User.create(username='huey', login_count=0)

# Simulate the user logging in. The login count and timestamp will be
# either created or updated correctly.
now = datetime.now()
rowid = (User
         .insert(username='huey', last_login=now, login_count=1)
             conflict_target=[User.username],  # Which constraint?
             preserve=[User.last_login],  # Use the value we would have inserted.
             update={User.login_count: User.login_count + 1})

In the above example, we could safely invoke the upsert query as many times as we wanted. The login count will be incremented atomically, the last login column will be updated, and no duplicate rows will be created.


The main difference between MySQL and Postgresql/SQLite is that Postgresql and SQLite require that you specify a conflict_target.

Here is a more advanced (if contrived) example using the EXCLUDED namespace. The EXCLUDED helper allows us to reference values in the conflicting data. For our example, we’ll assume a simple table mapping a unique key (string) to a value (integer):

class KV(Model):
    key = CharField(unique=True)
    value = IntegerField()

# Create one row.
KV.create(key='k1', value=1)

# Demonstrate usage of EXCLUDED.
# Here we will attempt to insert a new value for a given key. If that
# key already exists, then we will update its value with the *sum* of its
# original value and the value we attempted to insert -- provided that
# the new value is larger than the original value.
query = (KV.insert(key='k1', value=10)
                      update={KV.value: KV.value + EXCLUDED.value},
                      where=(EXCLUDED.value > KV.value)))

# Executing the above query will result in the following data being
# present in the "kv" table:
# (key='k1', value=11)

# If we attempted to execute the query *again*, then nothing would be
# updated, as the new value (10) is now less than the value in the
# original row (11).

There are several important concepts to understand when using ON CONFLICT:

  • conflict_target=: which column(s) have the UNIQUE constraint. For a user table, this might be the user’s email.

  • preserve=: if a conflict occurs, this parameter is used to indicate which values from the new data we wish to update.

  • update=: if a conflict occurs, this is a mapping of data to apply to the pre-existing row.

  • EXCLUDED: this “magic” namespace allows you to reference the new data that would have been inserted if the constraint hadn’t failed.

Full example:

class User(Model):
    email = CharField(unique=True)  # Unique identifier for user.
    last_login = DateTimeField()
    login_count = IntegerField(default=0)
    ip_log = TextField(default='')

# Demonstrates the above 4 concepts.
def login(email, ip):
    rowid = (User
             .insert({User.email: email,
                      User.last_login: datetime.now(),
                      User.login_count: 1,
                      User.ip_log: ip})
                 # If the INSERT fails due to a constraint violation on the
                 # user email, then perform an UPDATE instead.

                 # Set the "last_login" to the value we would have inserted
                 # (our call to datetime.now()).

                 # Increment the user's login count and prepend the new IP
                 # to the user's ip history.
                 update={User.login_count: User.login_count + 1,
                         User.ip_log: fn.CONCAT(EXCLUDED.ip_log, ',', User.ip_log)})

    return rowid

# This will insert the initial row, returning the new row id (1).
print(login('test@example.com', '127.1'))

# Because test@example.com exists, this will trigger the UPSERT. The row id
# from above is returned again (1).
print(login('test@example.com', '127.2'))

u = User.get()
print(u.login_count, u.ip_log)

# Prints "2 127.2,127.1"

For more information, see Insert.on_conflict() and OnConflict.

Deleting records

To delete a single model instance, you can use the Model.delete_instance() shortcut. delete_instance() will delete the given model instance and can optionally delete any dependent objects recursively (by specifying recursive=True).

>>> user = User.get(User.id == 1)
>>> user.delete_instance()  # Returns the number of rows deleted.

>>> User.get(User.id == 1)
UserDoesNotExist: instance matching query does not exist:
SQL: SELECT t1."id", t1."username" FROM "user" AS t1 WHERE t1."id" = ?

To delete an arbitrary set of rows, you can issue a DELETE query. The following will delete all Tweet objects that are over one year old:

>>> query = Tweet.delete().where(Tweet.creation_date < one_year_ago)
>>> query.execute()  # Returns the number of rows deleted.

For more information, see the documentation on:

Selecting a single record

You can use the Model.get() method to retrieve a single instance matching the given query. For primary-key lookups, you can also use the shortcut method Model.get_by_id().

This method is a shortcut that calls Model.select() with the given query, but limits the result set to a single row. Additionally, if no model matches the given query, a DoesNotExist exception will be raised.

>>> User.get(User.id == 1)
<__main__.User object at 0x25294d0>

>>> User.get_by_id(1)  # Same as above.
<__main__.User object at 0x252df10>

>>> User[1]  # Also same as above.
<__main__.User object at 0x252dd10>

>>> User.get(User.id == 1).username

>>> User.get(User.username == 'Charlie')
<__main__.User object at 0x2529410>

>>> User.get(User.username == 'nobody')
UserDoesNotExist: instance matching query does not exist:
SQL: SELECT t1."id", t1."username" FROM "user" AS t1 WHERE t1."username" = ?
PARAMS: ['nobody']

For more advanced operations, you can use SelectBase.get(). The following query retrieves the latest tweet from the user named charlie:

>>> (Tweet
...  .select()
...  .join(User)
...  .where(User.username == 'charlie')
...  .order_by(Tweet.created_date.desc())
...  .get())
<__main__.Tweet object at 0x2623410>

For more information, see the documentation on:

Create or get

Peewee has one helper method for performing “get/create” type operations: Model.get_or_create(), which first attempts to retrieve the matching row. Failing that, a new row will be created.

For “create or get” type logic, typically one would rely on a unique constraint or primary key to prevent the creation of duplicate objects. As an example, let’s say we wish to implement registering a new user account using the example User model. The User model has a unique constraint on the username field, so we will rely on the database’s integrity guarantees to ensure we don’t end up with duplicate usernames:

    with db.atomic():
        return User.create(username=username)
except peewee.IntegrityError:
    # `username` is a unique column, so this username already exists,
    # making it safe to call .get().
    return User.get(User.username == username)

You can easily encapsulate this type of logic as a classmethod on your own Model classes.

The above example first attempts at creation, then falls back to retrieval, relying on the database to enforce a unique constraint. If you prefer to attempt to retrieve the record first, you can use get_or_create(). This method is implemented along the same lines as the Django function of the same name. You can use the Django-style keyword argument filters to specify your WHERE conditions. The function returns a 2-tuple containing the instance and a boolean value indicating if the object was created.

Here is how you might implement user account creation using get_or_create():

user, created = User.get_or_create(username=username)

Suppose we have a different model Person and would like to get or create a person object. The only conditions we care about when retrieving the Person are their first and last names, but if we end up needing to create a new record, we will also specify their date-of-birth and favorite color:

person, created = Person.get_or_create(
    defaults={'dob': dob, 'favorite_color': 'green'})

Any keyword argument passed to get_or_create() will be used in the get() portion of the logic, except for the defaults dictionary, which will be used to populate values on newly-created instances.

For more details read the documentation for Model.get_or_create().

Selecting multiple records

We can use Model.select() to retrieve rows from the table. When you construct a SELECT query, the database will return any rows that correspond to your query. Peewee allows you to iterate over these rows, as well as use indexing and slicing operations:

>>> query = User.select()
>>> [user.username for user in query]
['Charlie', 'Huey', 'Peewee']

>>> query[1]
<__main__.User at 0x7f83e80f5550>

>>> query[1].username

>>> query[:2]
[<__main__.User at 0x7f83e80f53a8>, <__main__.User at 0x7f83e80f5550>]

Select queries are smart, in that you can iterate, index and slice the query multiple times but the query is only executed once.

In the following example, we will simply call select() and iterate over the return value, which is an instance of Select. This will return all the rows in the User table:

>>> for user in User.select():
...     print(user.username)


Subsequent iterations of the same query will not hit the database as the results are cached. To disable this behavior (to reduce memory usage), call Select.iterator() when iterating.

When iterating over a model that contains a foreign key, be careful with the way you access values on related models. Accidentally resolving a foreign key or iterating over a back-reference can cause N+1 query behavior.

When you create a foreign key, such as Tweet.user, you can use the backref to create a back-reference (User.tweets). Back-references are exposed as Select instances:

>>> tweet = Tweet.get()
>>> tweet.user  # Accessing a foreign key returns the related model.
<tw.User at 0x7f3ceb017f50>

>>> user = User.get()
>>> user.tweets  # Accessing a back-reference returns a query.
<peewee.ModelSelect at 0x7f73db3bafd0>

You can iterate over the user.tweets back-reference just like any other Select:

>>> for tweet in user.tweets:
...     print(tweet.message)
hello world
this is fun
look at this picture of my food

In addition to returning model instances, Select queries can return dictionaries, tuples and namedtuples. Depending on your use-case, you may find it easier to work with rows as dictionaries, for example:

>>> query = User.select().dicts()
>>> for row in query:
...     print(row)

{'id': 1, 'username': 'Charlie'}
{'id': 2, 'username': 'Huey'}
{'id': 3, 'username': 'Peewee'}

See namedtuples(), tuples(), dicts() for more information.

Iterating over large result-sets

By default peewee will cache the rows returned when iterating over a Select query. This is an optimization to allow multiple iterations as well as indexing and slicing without causing additional queries. This caching can be problematic, however, when you plan to iterate over a large number of rows.

To reduce the amount of memory used by peewee when iterating over a query, use the iterator() method. This method allows you to iterate without caching each model returned, using much less memory when iterating over large result sets.

# Let's assume we've got 10 million stat objects to dump to a csv file.
stats = Stat.select()

# Our imaginary serializer class
serializer = CSVSerializer()

# Loop over all the stats and serialize.
for stat in stats.iterator():

For simple queries you can see further speed improvements by returning rows as dictionaries, namedtuples or tuples. The following methods can be used on any Select query to change the result row type:

Don’t forget to append the iterator() method call to also reduce memory consumption. For example, the above code might look like:

# Let's assume we've got 10 million stat objects to dump to a csv file.
stats = Stat.select()

# Our imaginary serializer class
serializer = CSVSerializer()

# Loop over all the stats (rendered as tuples, without caching) and serialize.
for stat_tuple in stats.tuples().iterator():

When iterating over a large number of rows that contain columns from multiple tables, peewee will reconstruct the model graph for each row returned. This operation can be slow for complex graphs. For example, if we were selecting a list of tweets along with the username and avatar of the tweet’s author, Peewee would have to create two objects for each row (a tweet and a user). In addition to the above row-types, there is a fourth method objects() which will return the rows as model instances, but will not attempt to resolve the model graph.

For example:

query = (Tweet
         .select(Tweet, User)  # Select tweet and user data.

# Note that the user columns are stored in a separate User instance
# accessible at tweet.user:
for tweet in query:
    print(tweet.user.username, tweet.content)

# Using ".objects()" will not create the tweet.user object and assigns all
# user attributes to the tweet instance:
for tweet in query.objects():
    print(tweet.username, tweet.content)

For maximum performance, you can execute queries and then iterate over the results using the underlying database cursor. Database.execute() accepts a query object, executes the query, and returns a DB-API 2.0 Cursor object. The cursor will return the raw row-tuples:

query = Tweet.select(Tweet.content, User.username).join(User)
cursor = database.execute(query)
for (content, username) in cursor:
    print(username, '->', content)

Filtering records

You can filter for particular records using normal python operators. Peewee supports a wide variety of query operators.

>>> user = User.get(User.username == 'Charlie')
>>> for tweet in Tweet.select().where(Tweet.user == user, Tweet.is_published == True):
...     print(tweet.user.username, '->', tweet.message)
Charlie -> hello world
Charlie -> this is fun

>>> for tweet in Tweet.select().where(Tweet.created_date < datetime.datetime(2011, 1, 1)):
...     print(tweet.message, tweet.created_date)
Really old tweet 2010-01-01 00:00:00

You can also filter across joins:

>>> for tweet in Tweet.select().join(User).where(User.username == 'Charlie'):
...     print(tweet.message)
hello world
this is fun
look at this picture of my food

If you want to express a complex query, use parentheses and python’s bitwise or and and operators:

>>> Tweet.select().join(User).where(
...     (User.username == 'Charlie') |
...     (User.username == 'Peewee Herman'))


Note that Peewee uses bitwise operators (& and |) rather than logical operators (and and or). The reason for this is that Python coerces the return value of logical operations to a boolean value. This is also the reason why “IN” queries must be expressed using .in_() rather than the in operator.

Check out the table of query operations to see what types of queries are possible.


A lot of fun things can go in the where clause of a query, such as:

  • A field expression, e.g. User.username == 'Charlie'

  • A function expression, e.g. fn.Lower(fn.Substr(User.username, 1, 1)) == 'a'

  • A comparison of one column to another, e.g. Employee.salary < (Employee.tenure * 1000) + 40000

You can also nest queries, for example tweets by users whose username starts with “a”:

# get users whose username starts with "a"
a_users = User.select().where(fn.Lower(fn.Substr(User.username, 1, 1)) == 'a')

# the ".in_()" method signifies an "IN" query
a_user_tweets = Tweet.select().where(Tweet.user.in_(a_users))

More query examples


For a wide range of example queries, see the Query Examples document, which shows how to implements queries from the PostgreSQL Exercises website.

Get active users:

User.select().where(User.active == True)

Get users who are either staff or superusers:

    (User.is_staff == True) | (User.is_superuser == True))

Get tweets by user named “charlie”:

Tweet.select().join(User).where(User.username == 'charlie')

Get tweets by staff or superusers (assumes FK relationship):

    (User.is_staff == True) | (User.is_superuser == True))

Get tweets by staff or superusers using a subquery:

staff_super = User.select(User.id).where(
    (User.is_staff == True) | (User.is_superuser == True))

Sorting records

To return rows in order, use the order_by() method:

>>> for t in Tweet.select().order_by(Tweet.created_date):
...     print(t.pub_date)
2010-01-01 00:00:00
2011-06-07 14:08:48
2011-06-07 14:12:57

>>> for t in Tweet.select().order_by(Tweet.created_date.desc()):
...     print(t.pub_date)
2011-06-07 14:12:57
2011-06-07 14:08:48
2010-01-01 00:00:00

You can also use + and - prefix operators to indicate ordering:

# The following queries are equivalent:

Tweet.select().order_by(-Tweet.created_date)  # Note the "-" prefix.

# Similarly you can use "+" to indicate ascending order, though ascending
# is the default when no ordering is otherwise specified.

You can also order across joins. Assuming you want to order tweets by the username of the author, then by created_date:

query = (Tweet
         .order_by(User.username, Tweet.created_date.desc()))
SELECT t1."id", t1."user_id", t1."message", t1."is_published", t1."created_date"
FROM "tweet" AS t1
INNER JOIN "user" AS t2
  ON t1."user_id" = t2."id"
ORDER BY t2."username", t1."created_date" DESC

When sorting on a calculated value, you can either include the necessary SQL expressions, or reference the alias assigned to the value. Here are two examples illustrating these methods:

# Let's start with our base query. We want to get all usernames and the number of
# tweets they've made. We wish to sort this list from users with most tweets to
# users with fewest tweets.
query = (User
         .select(User.username, fn.COUNT(Tweet.id).alias('num_tweets'))
         .join(Tweet, JOIN.LEFT_OUTER)

You can order using the same COUNT expression used in the select clause. In the example below we are ordering by the COUNT() of tweet ids descending:

query = (User
         .select(User.username, fn.COUNT(Tweet.id).alias('num_tweets'))
         .join(Tweet, JOIN.LEFT_OUTER)

Alternatively, you can reference the alias assigned to the calculated value in the select clause. This method has the benefit of being a bit easier to read. Note that we are not referring to the named alias directly, but are wrapping it using the SQL helper:

query = (User
         .select(User.username, fn.COUNT(Tweet.id).alias('num_tweets'))
         .join(Tweet, JOIN.LEFT_OUTER)

Or, to do things the “peewee” way:

ntweets = fn.COUNT(Tweet.id)
query = (User
         .select(User.username, ntweets.alias('num_tweets'))
         .join(Tweet, JOIN.LEFT_OUTER)

Getting random records

Occasionally you may want to pull a random record from the database. You can accomplish this by ordering by the random or rand function (depending on your database):

Postgresql and Sqlite use the Random function:

# Pick 5 lucky winners:

MySQL uses Rand:

# Pick 5 lucky winners:

Paginating records

The paginate() method makes it easy to grab a page or records. paginate() takes two parameters, page_number, and items_per_page.


Page numbers are 1-based, so the first page of results will be page 1.

>>> for tweet in Tweet.select().order_by(Tweet.id).paginate(2, 10):
...     print(tweet.message)
tweet 10
tweet 11
tweet 12
tweet 13
tweet 14
tweet 15
tweet 16
tweet 17
tweet 18
tweet 19

If you would like more granular control, you can always use limit() and offset().

Counting records

You can count the number of rows in any select query:

>>> Tweet.select().count()
>>> Tweet.select().where(Tweet.id > 50).count()

Peewee will wrap your query in an outer query that performs a count, which results in SQL like:

SELECT COUNT(1) FROM ( ... your query ... );

Aggregating records

Suppose you have some users and want to get a list of them along with the count of tweets in each.

query = (User
         .select(User, fn.Count(Tweet.id).alias('count'))
         .join(Tweet, JOIN.LEFT_OUTER)

The resulting query will return User objects with all their normal attributes plus an additional attribute count which will contain the count of tweets for each user. We use a left outer join to include users who have no tweets.

Let’s assume you have a tagging application and want to find tags that have a certain number of related objects. For this example we’ll use some different models in a many-to-many configuration:

class Photo(Model):
    image = CharField()

class Tag(Model):
    name = CharField()

class PhotoTag(Model):
    photo = ForeignKeyField(Photo)
    tag = ForeignKeyField(Tag)

Now say we want to find tags that have at least 5 photos associated with them:

query = (Tag
         .having(fn.Count(Photo.id) > 5))

This query is equivalent to the following SQL:

SELECT t1."id", t1."name"
FROM "tag" AS t1
INNER JOIN "phototag" AS t2 ON t1."id" = t2."tag_id"
INNER JOIN "photo" AS t3 ON t2."photo_id" = t3."id"
GROUP BY t1."id", t1."name"
HAVING Count(t3."id") > 5

Suppose we want to grab the associated count and store it on the tag:

query = (Tag
         .select(Tag, fn.Count(Photo.id).alias('count'))
         .having(fn.Count(Photo.id) > 5))

Retrieving Scalar Values

You can retrieve scalar values by calling Query.scalar(). For instance:

>>> PageView.select(fn.Count(fn.Distinct(PageView.url))).scalar()

You can retrieve multiple scalar values by passing as_tuple=True:

>>> Employee.select(
...     fn.Min(Employee.salary), fn.Max(Employee.salary)
... ).scalar(as_tuple=True)
(30000, 50000)

Window functions

A Window function refers to an aggregate function that operates on a sliding window of data that is being processed as part of a SELECT query. Window functions make it possible to do things like:

  1. Perform aggregations against subsets of a result-set.

  2. Calculate a running total.

  3. Rank results.

  4. Compare a row value to a value in the preceding (or succeeding!) row(s).

peewee comes with support for SQL window functions, which can be created by calling Function.over() and passing in your partitioning or ordering parameters.

For the following examples, we’ll use the following model and sample data:

class Sample(Model):
    counter = IntegerField()
    value = FloatField()

data = [(1, 10),
        (1, 20),
        (2, 1),
        (2, 3),
        (3, 100)]
Sample.insert_many(data, fields=[Sample.counter, Sample.value]).execute()

Our sample table now contains:



















Ordered Windows

Let’s calculate a running sum of the value field. In order for it to be a “running” sum, we need it to be ordered, so we’ll order with respect to the Sample’s id field:

query = Sample.select(

for sample in query:
    print(sample.counter, sample.value, sample.total)

# 1    10.    10.
# 1    20.    30.
# 2     1.    31.
# 2     3.    34.
# 3   100    134.

For another example, we’ll calculate the difference between the current value and the previous value, when ordered by the id:

difference = Sample.value - fn.LAG(Sample.value, 1).over(order_by=[Sample.id])
query = Sample.select(

for sample in query:
    print(sample.counter, sample.value, sample.diff)

# 1    10.   NULL
# 1    20.    10.  -- (20 - 10)
# 2     1.   -19.  -- (1 - 20)
# 2     3.     2.  -- (3 - 1)
# 3   100     97.  -- (100 - 3)

Partitioned Windows

Let’s calculate the average value for each distinct “counter” value. Notice that there are three possible values for the counter field (1, 2, and 3). We can do this by calculating the AVG() of the value column over a window that is partitioned depending on the counter field:

query = Sample.select(

for sample in query:
    print(sample.counter, sample.value, sample.cavg)

# 1    10.    15.
# 1    20.    15.
# 2     1.     2.
# 2     3.     2.
# 3   100    100.

We can use ordering within partitions by specifying both the order_by and partition_by parameters. For an example, let’s rank the samples by value within each distinct counter group.

query = Sample.select(

for sample in query:
    print(sample.counter, sample.value, sample.rank)

# 1    10.    1
# 1    20.    2
# 2     1.    1
# 2     3.    2
# 3   100     1

Bounded windows

By default, window functions are evaluated using an unbounded preceding start for the window, and the current row as the end. We can change the bounds of the window our aggregate functions operate on by specifying a start and/or end in the call to Function.over(). Additionally, Peewee comes with helper-methods on the Window object for generating the appropriate boundary references:

  • Window.CURRENT_ROW - attribute that references the current row.

  • Window.preceding() - specify number of row(s) preceding, or omit number to indicate all preceding rows.

  • Window.following() - specify number of row(s) following, or omit number to indicate all following rows.

To examine how boundaries work, we’ll calculate a running total of the value column, ordered with respect to id, but we’ll only look the running total of the current row and it’s two preceding rows:

query = Sample.select(

for sample in query:
    print(sample.counter, sample.value, sample.rsum)

# 1    10.    10.
# 1    20.    30.  -- (20 + 10)
# 2     1.    31.  -- (1 + 20 + 10)
# 2     3.    24.  -- (3 + 1 + 20)
# 3   100    104.  -- (100 + 3 + 1)


Technically we did not need to specify the end=Window.CURRENT because that is the default. It was shown in the example for demonstration.

Let’s look at another example. In this example we will calculate the “opposite” of a running total, in which the total sum of all values is decreased by the value of the samples, ordered by id. To accomplish this, we’ll calculate the sum from the current row to the last row.

query = Sample.select(

# 1    10.   134.  -- (10 + 20 + 1 + 3 + 100)
# 1    20.   124.  -- (20 + 1 + 3 + 100)
# 2     1.   104.  -- (1 + 3 + 100)
# 2     3.   103.  -- (3 + 100)
# 3   100    100.  -- (100)

Filtered Aggregates

Aggregate functions may also support filter functions (Postgres and Sqlite 3.25+), which get translated into a FILTER (WHERE...) clause. Filter expressions are added to an aggregate function with the Function.filter() method.

For an example, we will calculate the running sum of the value field with respect to the id, but we will filter-out any samples whose counter=2.

query = Sample.select(
    fn.SUM(Sample.value).filter(Sample.counter != 2).over(

for sample in query:
    print(sample.counter, sample.value, sample.csum)

# 1    10.    10.
# 1    20.    30.
# 2     1.    30.
# 2     3.    30.
# 3   100    130.


The call to filter() must precede the call to over().

Reusing Window Definitions

If you intend to use the same window definition for multiple aggregates, you can create a Window object. The Window object takes the same parameters as Function.over(), and can be passed to the over() method in-place of the individual parameters.

Here we’ll declare a single window, ordered with respect to the sample id, and call several window functions using that window definition:

win = Window(order_by=[Sample.id])
query = Sample.select(
).window(win)  # Include our window definition in query.

for row in query.tuples():

# counter  value  lead()  lag()  sum()
# 1          10.     20.   NULL    10.
# 1          20.      1.    10.    30.
# 2           1.      3.    20.    31.
# 2           3.    100.     1.    34.
# 3         100.    NULL     3.   134.

Multiple window definitions

In the previous example, we saw how to declare a Window definition and re-use it for multiple different aggregations. You can include as many window definitions as you need in your queries, but it is necessary to ensure each window has a unique alias:

w1 = Window(order_by=[Sample.id]).alias('w1')
w2 = Window(partition_by=[Sample.counter]).alias('w2')
query = Sample.select(
    fn.SUM(Sample.value).over(w1).alias('rsum'),  # Running total.
    fn.AVG(Sample.value).over(w2).alias('cavg')   # Avg per category.
).window(w1, w2)  # Include our window definitions.

for sample in query:
    print(sample.counter, sample.value, sample.rsum, sample.cavg)

# counter  value   rsum     cavg
# 1          10.     10.     15.
# 1          20.     30.     15.
# 2           1.     31.      2.
# 2           3.     34.      2.
# 3         100     134.    100.

Similarly, if you have multiple window definitions that share similar definitions, it is possible to extend a previously-defined window definition. For example, here we will be partitioning the data-set by the counter value, so we’ll be doing our aggregations with respect to the counter. Then we’ll define a second window that extends this partitioning, and adds an ordering clause:

w1 = Window(partition_by=[Sample.counter]).alias('w1')

# By extending w1, this window definition will also be partitioned
# by "counter".
w2 = Window(extends=w1, order_by=[Sample.value.desc()]).alias('w2')

query = (Sample
         .select(Sample.counter, Sample.value,
         .window(w1, w2)

for sample in query:
    print(sample.counter, sample.value, sample.group_sum, sample.revrank)

# counter  value   group_sum   revrank
# 1        10.     30.         2
# 1        20.     30.         1
# 2        1.      4.          2
# 2        3.      4.          1
# 3        100.    100.        1

Frame types: RANGE vs ROWS vs GROUPS

Depending on the frame type, the database will process ordered groups differently. Let’s create two additional Sample rows to visualize the difference:

>>> Sample.create(counter=1, value=20.)
<Sample 6>
>>> Sample.create(counter=2, value=1.)
<Sample 7>

Our table now contains:

























Let’s examine the difference by calculating a “running sum” of the samples, ordered with respect to the counter and value fields. To specify the frame type, we can use either:

The behavior of RANGE, when there are logical duplicates, may lead to unexpected results:

query = Sample.select(
        order_by=[Sample.counter, Sample.value],

for sample in query.order_by(Sample.counter, Sample.value):
    print(sample.counter, sample.value, sample.rsum)

# counter  value   rsum
# 1          10.     10.
# 1          20.     50.
# 1          20.     50.
# 2           1.     52.
# 2           1.     52.
# 2           3.     55.
# 3         100     155.

With the inclusion of the new rows we now have some rows that have duplicate category and value values. The RANGE frame type causes these duplicates to be evaluated together rather than separately.

The more expected result can be achieved by using ROWS as the frame-type:

query = Sample.select(
        order_by=[Sample.counter, Sample.value],

for sample in query.order_by(Sample.counter, Sample.value):
    print(sample.counter, sample.value, sample.rsum)

# counter  value   rsum
# 1          10.     10.
# 1          20.     30.
# 1          20.     50.
# 2           1.     51.
# 2           1.     52.
# 2           3.     55.
# 3         100     155.

Peewee uses these rules for determining what frame-type to use:

  • If the user specifies a frame_type, that frame type will be used.

  • If start and/or end boundaries are specified Peewee will default to using ROWS.

  • If the user did not specify frame type or start/end boundaries, Peewee will use the database default, which is RANGE.

The Window.GROUPS frame type looks at the window range specification in terms of groups of rows, based on the ordering term(s). Using GROUPS, we can define the frame so it covers distinct groupings of rows. Let’s look at an example:

query = (Sample
         .select(Sample.counter, Sample.value,
                    order_by=[Sample.counter, Sample.value],
         .order_by(Sample.counter, Sample.value))

for sample in query:
    print(sample.counter, sample.value, sample.gsum)

#  counter   value    gsum
#  1         10       10
#  1         20       50
#  1         20       50   (10) + (20+0)
#  2         1        42
#  2         1        42   (20+20) + (1+1)
#  2         3        5    (1+1) + 3
#  3         100      103  (3) + 100

As you can hopefully infer, the window is grouped by its ordering term, which is (counter, value). We are looking at a window that extends between one previous group and the current group.


For information about the window function APIs, see:

For general information on window functions, read the postgres window functions tutorial

Additionally, the postgres docs and the sqlite docs contain a lot of good information.

Retrieving row tuples / dictionaries / namedtuples

Sometimes you do not need the overhead of creating model instances and simply want to iterate over the row data without needing all the APIs provided Model. To do this, use:

stats = (Stat
         .select(Stat.url, fn.Count(Stat.url))

# iterate over a list of 2-tuples containing the url and count
for stat_url, stat_count in stats:
    print(stat_url, stat_count)

Similarly, you can return the rows from the cursor as dictionaries using dicts():

stats = (Stat
         .select(Stat.url, fn.Count(Stat.url).alias('ct'))

# iterate over a list of 2-tuples containing the url and count
for stat in stats:
    print(stat['url'], stat['ct'])

Returning Clause

PostgresqlDatabase supports a RETURNING clause on UPDATE, INSERT and DELETE queries. Specifying a RETURNING clause allows you to iterate over the rows accessed by the query.

By default, the return values upon execution of the different queries are:

  • INSERT - auto-incrementing primary key value of the newly-inserted row. When not using an auto-incrementing primary key, Postgres will return the new row’s primary key, but SQLite and MySQL will not.

  • UPDATE - number of rows modified

  • DELETE - number of rows deleted

When a returning clause is used the return value upon executing a query will be an iterable cursor object.

Postgresql allows, via the RETURNING clause, to return data from the rows inserted or modified by a query.

For example, let’s say you have an Update that deactivates all user accounts whose registration has expired. After deactivating them, you want to send each user an email letting them know their account was deactivated. Rather than writing two queries, a SELECT and an UPDATE, you can do this in a single UPDATE query with a RETURNING clause:

query = (User
         .where(User.registration_expired == True)

# Send an email to every user that was deactivated.
for deactivate_user in query.execute():

The RETURNING clause is also available on Insert and Delete. When used with INSERT, the newly-created rows will be returned. When used with DELETE, the deleted rows will be returned.

The only limitation of the RETURNING clause is that it can only consist of columns from tables listed in the query’s FROM clause. To select all columns from a particular table, you can simply pass in the Model class.

As another example, let’s add a user and set their creation-date to the server-generated current timestamp. We’ll create and retrieve the new user’s ID, Email and the creation timestamp in a single query:

query = (User
         .insert(email='foo@bar.com', created=fn.now())
         .returning(User))  # Shorthand for all columns on User.

# When using RETURNING, execute() returns a cursor.
cursor = query.execute()

# Get the user object we just inserted and log the data:
user = cursor[0]
logger.info('Created user %s (id=%s) at %s', user.email, user.id, user.created)

By default the cursor will return Model instances, but you can specify a different row type:

data = [{'name': 'charlie'}, {'name': 'huey'}, {'name': 'mickey'}]
query = (User
         .returning(User.id, User.username)

for new_user in query.execute():
    print('Added user "%s", id=%s' % (new_user['username'], new_user['id']))

Just as with Select queries, you can specify various result row types.

Common Table Expressions

Peewee supports the inclusion of common table expressions (CTEs) in all types of queries. CTEs may be useful for:

  • Factoring out a common subquery.

  • Grouping or filtering by a column derived in the CTE’s result set.

  • Writing recursive queries.

To declare a Select query for use as a CTE, use cte() method, which wraps the query in a CTE object. To indicate that a CTE should be included as part of a query, use the Query.with_cte() method, passing a list of CTE objects.

Simple Example

For an example, let’s say we have some data points that consist of a key and a floating-point value. Let’s define our model and populate some test data:

class Sample(Model):
    key = TextField()
    value = FloatField()

data = (
    ('a', (1.25, 1.5, 1.75)),
    ('b', (2.1, 2.3, 2.5, 2.7, 2.9)),
    ('c', (3.5, 3.5)))

# Populate data.
for key, values in data:
    Sample.insert_many([(key, value) for value in values],
                       fields=[Sample.key, Sample.value]).execute()

Let’s use a CTE to calculate, for each distinct key, which values were above-average for that key.

# First we'll declare the query that will be used as a CTE. This query
# simply determines the average value for each key.
cte = (Sample
       .select(Sample.key, fn.AVG(Sample.value).alias('avg_value'))
       .cte('key_avgs', columns=('key', 'avg_value')))

# Now we'll query the sample table, using our CTE to find rows whose value
# exceeds the average for the given key. We'll calculate how far above the
# average the given sample's value is, as well.
query = (Sample
         .select(Sample.key, Sample.value)
         .join(cte, on=(Sample.key == cte.c.key))
         .where(Sample.value > cte.c.avg_value)

We can iterate over the samples returned by the query to see which samples had above-average values for their given group:

>>> for sample in query:
...     print(sample.key, sample.value)

# 'a', 1.75
# 'b', 2.7
# 'b', 2.9

Complex Example

For a more complete example, let’s consider the following query which uses multiple CTEs to find per-product sales totals in only the top sales regions. Our model looks like this:

class Order(Model):
    region = TextField()
    amount = FloatField()
    product = TextField()
    quantity = IntegerField()

Here is how the query might be written in SQL. This example can be found in the postgresql documentation.

WITH regional_sales AS (
    SELECT region, SUM(amount) AS total_sales
    FROM orders
    GROUP BY region
  ), top_regions AS (
    SELECT region
    FROM regional_sales
    WHERE total_sales > (SELECT SUM(total_sales) / 10 FROM regional_sales)
SELECT region,
       SUM(quantity) AS product_units,
       SUM(amount) AS product_sales
FROM orders
WHERE region IN (SELECT region FROM top_regions)
GROUP BY region, product;

With Peewee, we would write:

reg_sales = (Order

top_regions = (reg_sales
               .where(reg_sales.c.total_sales > (
                   reg_sales.select(fn.SUM(reg_sales.c.total_sales) / 10)))

query = (Order
         .group_by(Order.region, Order.product)
         .with_cte(reg_sales, top_regions))

Recursive CTEs

Peewee supports recursive CTEs. Recursive CTEs can be useful when, for example, you have a tree data-structure represented by a parent-link foreign key. Suppose, for example, that we have a hierarchy of categories for an online bookstore. We wish to generate a table showing all categories and their absolute depths, along with the path from the root to the category.

We’ll assume the following model definition, in which each category has a foreign-key to its immediate parent category:

class Category(Model):
    name = TextField()
    parent = ForeignKeyField('self', backref='children', null=True)

To list all categories along with their depth and parents, we can use a recursive CTE:

# Define the base case of our recursive CTE. This will be categories that
# have a null parent foreign-key.
Base = Category.alias()
level = Value(1).alias('level')
path = Base.name.alias('path')
base_case = (Base
             .select(Base.id, Base.name, Base.parent, level, path)
             .cte('base', recursive=True))

# Define the recursive terms.
RTerm = Category.alias()
rlevel = (base_case.c.level + 1).alias('level')
rpath = base_case.c.path.concat('->').concat(RTerm.name).alias('path')
recursive = (RTerm
             .select(RTerm.id, RTerm.name, RTerm.parent, rlevel, rpath)
             .join(base_case, on=(RTerm.parent == base_case.c.id)))

# The recursive CTE is created by taking the base case and UNION ALL with
# the recursive term.
cte = base_case.union_all(recursive)

# We will now query from the CTE to get the categories, their levels,  and
# their paths.
query = (cte
         .select_from(cte.c.name, cte.c.level, cte.c.path)

# We can now iterate over a list of all categories and print their names,
# absolute levels, and path from root -> category.
for category in query:
    print(category.name, category.level, category.path)

# Example output:
# root, 1, root
# p1, 2, root->p1
# c1-1, 3, root->p1->c1-1
# c1-2, 3, root->p1->c1-2
# p2, 2, root->p2
# c2-1, 3, root->p2->c2-1

Data-Modifying CTE

Peewee supports data-modifying CTE’s.

Example of using a data-modifying CTE to move data from one table to an archive table, using a single query:

class Event(Model):
    name = CharField()
    timestamp = DateTimeField()

class Archive(Model):
    name = CharField()
    timestamp = DateTimeField()

# Move rows older than 24 hours from the Event table to the Archive.
cte = (Event
       .where(Event.timestamp < (datetime.now() - timedelta(days=1)))

# Create a simple SELECT to get the resulting rows from the CTE.
src = Select((cte,), (cte.c.id, cte.c.name, cte.c.timestamp))

# Insert into the archive table whatever data was returned by the DELETE.
res = (Archive
       .insert_from(src, (Archive.id, Archive.name, Archive.timestamp))

The above corresponds to, roughly, the following SQL:

WITH "moved_rows" AS (
    DELETE FROM "event" WHERE ("timestamp" < XXXX-XX-XXTXX:XX:XX)
    RETURNING "id", "name", "timestamp")
INSERT INTO "archive" ("id", "name", "timestamp")
SELECT "moved_rows"."id", "moved_rows"."name", "moved_rows"."timestamp"
FROM "moved_rows";

For additional examples, refer to the tests in models.py and sql.py:

Foreign Keys and Joins

This section has been moved into its own document: Relationships and Joins.